SAL-Net: Self-Supervised Attribute Learning for Object Recognition and Segmentation

Table 1

Effectiveness of the proposed strategies, including multitask learning, features from multiple CNN layers, and joint learning with attributes.

Model	Strategies			CUB-2011 [10]		Pascal VOC [28]
	Multitask	Multifeature	Attribute learning	Acc (%)	aIoU (%)	Acc (%)	aIoU (%)
ST-cls				64.67	—	67.46	—
ST-seg				—	79.76	—	73.81
MT-SF	✓			70.37	74.22	69.11	74.02
MT-MF	✓	✓		71.45	82.61	71.42	77.59
AFE-Net [27]	✓	✓	✓	74.37	85.07	—	—
SAL-Net (ours)	✓	✓	✓	77.55	86.66	75.00	78.19